Actionability is read against the common region, not reason-relative ones

agent-decided, reversible. Individually actionable controls do not compose when each reason's gain is scored against that reason's own admissible region: an exhibited pair with jointly satisfiable demands, each meeting its margin exactly, has a per-reason gain of -1/8 after aggregation, because one reason's region does not constrain the coordinates the other moves. Scored against the intersection of the regions the composition is immediate — an infimum over a smaller set is larger, and an infimum is superadditive — with no convexity, separability or noninterference hypothesis.

The joint response correspondence therefore drops the reason-indexed comparison coordinate the dispatch proposed and keeps (w, u) with u = sum_r w^r zeta^r, standing beside one datum: the intersection is nonempty. So typed it is the image of a nonnegative orthant under a linear map intersected with the control set, which is convex whenever the control set is — the convexity the existence theory wants is what the minimal type already is, rather than an added assumption.

Rejected alternative: a joint response correspondence carrying per-reason scored challenges, which is the shape the interference countermodel exploits.